I have a RTX 3060 and RX 6600 XT. This behavior is consistent in Windows and Linux (ROCm). With the 3060, I can run a CUDA background app that uses 6GB of video RAM, have a -500 core underclock, run WoW at 1080p at max settings with RT shadows maxed even, and maintain 60 FPS no problem. The CUDA app slows down when games are running, which makes me think the driver is doing prioritization. I find this ideal as I don't have to close the CUDA app out with quick gaming sessions. Meanwhile, similar OpenCL apps on the 6600 perform slower overall, and game performance is bad while the OpenCL app is running regardless of GPU clocks (under, factory, overclock, max power limit, no downclocking, etc). If I want to game at all, I have to shut down the OpenCL app. 1. Does NVIDIA or AMD have anything explaining prioritization of compute vs 3D apps? 2. Can this prioritization on AMD be changed? I want OpenCL apps to lower performance in order for 3D performance to be good. 3. Why do similar compute apps perform notably better on the 3060? The difference on numbers is that the 6600 gets around 31 no matter the situation, whereas the 3060 does 35 non-ideally, and up to 50 ideally. The 6600 as I understand has better raw-performance. Basically, doing the same compute task on the 3060 has a notably better experience, and I want to know technical reasons why and if AMD can be configured to match it.
that is like asking why nvidia handles opengl better then amd, the answer is the same in both cases better optimizing/features or lack there of it on amd side atlest last time i checked anyway
Because CUDA cores are rarely used by games. They are separate hardware. Where as on AMD hardware the same part of the GPU that accelerates OpenCL is also used by games?
ampere are compute-oriented cards like gcn were but with good software to follow 3060 has nearly 13tf fp32 while 6600 has 9
Yeah if I wanted a vague answer I wouldn't have bothered asking Are there any resources or can you explain this further?
you need to be really specific in asking, besides every other answer you got is just as vague Now if you asked why OpenGL runs so much better on nvidia side vs amd you probably get way more detail, cause that is well documented
I didn't consider VRAM usage prior to this and it's likely I was running out of VRAM on the 6600. The OpenCL task on the 6600 (8G) is using about 5GB of VRAM and would leave about 3GB left for anything else (nowhere near enough for D2R). With WoW and the CUDA task on the 3060 (12G), I was around 9GB VRAM usage.
just as a follow up,here's a test ampere vs rdna 2 in workstation use 3080ti absolutely smashes 6900xt,that's not just raw power but better support for CUDA https://www.purepc.pl/akceleracja-s...deo-test-wydajnosci-kart-graficznych?page=0,6
Theoretically the 6xxxx series are really good for compute. But I do feel AMD hasn't really done work regarding putting that to use. Since for all intents and purposes I would think 6900 xt should be faster then 5950x in compute...